Fig 1.
Second and third administrative divisions on Bioko Island.
The thick black lines demarcate the four districts. Malabo and Baney make Bioko North and Luba and Riaba, Bioko South. Green areas are uninhabited nature reserves.
Table 1.
Basic summary statistics for all the maps, including the grid size, total population, maximum population density at 1 km2, the percentage of pixels that were empty or urban (with >1000 people per km2), and the Pareto Number, defined as the percentage x that holds a percentage 1 − x of the population.
Table 2.
Let a threshold, τ, define a categorization of population density.
In a gold standard map, G, a pixel is in the category if it is above the threshold: x ∈ Gτ if and only if x > τ. Otherwise, x ∉ Gτ. Similarly, the categorization is applied to a candidate map, M. Pixels are classified as true positives (TP), true negatives (TN), false negatives (FN), and false positives (FP) as described in the table. Accuracy profiles are plotted in Fig 6.
Fig 2.
Bioko population rendered at 1x1 km resolution.
Population colors range from greater than 10,000 persons per km2 to less than 10 per km2. Grey pixels represent uninhabited areas (population = 0).
Fig 3.
Gridded maps at their native resolution, zoomed into the Malabo area as visualization of a highly populated area for comparison.
Grey pixels represent uninhabited areas.
Fig 4.
These are scatter plots of the population count in each pixel as a function of the population count in each pixel according to BIMEP.
The solid black line indicates one-to-one agreement. The HRSL, WP-C, and WP-U maps are also shown aggregated to the 1 km scale. These maps have a background of grey points, which are the 1 km density of each pixel of these three maps. The R2 values of the 1 km maps are 51% for LS, 70% for HRSL, 62% for WP-C, -6% for WP-U, and -11% for GPW. At finer scales, where registration can be more difficult, the R2 values are -4% for HRSL, 48% for WP-C, and -2% for WP-U.
Fig 5.
A comparison of the distributions by land area and population density.
Solid lines are 1x1 km maps, and dashed lines are 100x100 m maps. A) To show how the population is distributed, we plotted the empirical cumulative distribution functions (eCDFs) of population density by land area; B) To show how the population is aggregated, we plotted the eCDFs by log population density. C) The population density binned by powers of 1.2.
Table 3.
This compares the goodness-of-fit ratio across the three maps, aggregating HRSL and both WP surfaces to 1 km resolution to match LS and GPW.
Normalization discounts the effect of uniform changes in population size, which provides a better comparison between high-and-low population districts.
Fig 6.
A) The proportion of the population in density categories defined by breakpoints of 1, 50, 250, and 1,000 people. B) The accuracy profile; C) The recall profile; D) The precision profile.
Table 4.
The accuracy, recall, and precision for the population classifications shown in the header and illustrated in Fig 6.
Fig 7.
Fraction of the population according to PfPR, expressed as cumulative distribution (A) and probability density functions (B).